Super-pixel high-frequency enhanced lightweight image super-resolution reconstruction method, device, equipment and medium

Abstract

The invention discloses a super-pixel high-frequency enhanced lightweight image super-resolution reconstruction method, device and equipment and a medium. Performing shallow feature extraction based on a convolutional network on an input image, grouping adjacent pixels into superpixels through local clustering, and clustering the superpixels according to similarity to obtain a local area; an intra-pixel attention module and an inter-pixel attention module are introduced, and the two attention mechanisms cooperate with each other in local and global feature extraction. Meanwhile, high-frequency prior information is introduced by using a high-frequency enhancement branch to capture high-frequency details, and the quality of a reconstructed image is improved. According to the method, pixels with similar characteristics are aggregated to form superpixels, and local and global information interaction is realized through an in-pixel attention module and an inter-pixel attention module respectively. Meanwhile, high-frequency prior information is added into the network by using a high-frequency enhancement module, so that the feature extraction capability of the network is effectively improved. A large number of experiments on a standard data set verify that the method has a better reconstruction effect.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and specifically, is a lightweight image super-resolution reconstruction method, device, equipment and medium with super-pixel high-frequency enhancement. Background Art

[0002] Image super-resolution (SR) aims to reconstruct a high-resolution (HR) image from its degraded low-resolution (LR) image. This method enables the generated image to have high visual quality and good texture details. Super-resolution is crucial for many devices, such as high-definition televisions, computer monitors, cameras, smartphones, and tablets. In addition, it is also used in various computer vision fields, such as object detection, medical imaging, security and surveillance imaging, face recognition, remote sensing images, etc.

[0003] SR reconstruction algorithms mainly include interpolation-based, reconstruction-based and learning-based methods. In recent years, due to the powerful feature expression ability of convolutional neural networks, deep learning-based methods have been widely used in SR tasks. Dong et al. first applied convolutional neural networks (CNN) to the field of single image super-resolution reconstruction and proposed the SRCNN network. However, due to the simple structure of SRCNN, the reconstruction performance is subject to certain limitations. Kim et al. proposed a deep super-resolution reconstruction network (Super-resolution using Very Deep Convolution Networks, VDSR), which increases the number of network layers through residual learning and solves the gradient vanishing phenomenon caused by increasing depth. Subsequently, in order to reduce the complexity of the model, Ahn et al. introduced the cascading residual network (CARN), which achieves more efficient super-resolution reconstruction on the residual network by using cascade connection operations.

[0004] The attention mechanism calculates different weights of feature vectors according to the input features to represent the different importance of features. It has been widely used in computer vision tasks such as image segmentation and target detection. Hu et al. proposed the Squeeze-and-Excitation (SE) module to learn the relationship between channels and assign different weights to channels according to the amount of information contained in the channels. Subsequently, in CBAM, channel attention and spatial attention were combined to achieve adaptive feature refinement. Guo et al. proposed the Large Kernel Attention (LKA) mechanism, which uses large-size convolution kernels to better represent the global structural features of the image and increase the receptive field. At the same time, the attention mechanism is also widely used in low-level visual tasks such as image super-resolution reconstruction. Zhang et al. proposed the residual channel attention module, which integrates channel attention into the residual module, fully utilizes the dependency between features, and achieves significant performance improvement. However, these works mainly focus on designing network structures to obtain global information, while ignoring the importance of incorporating high-frequency priors. In addition, the attention mechanism has high computational complexity and memory consumption. It is necessary to divide large images into small blocks for separate processing. The division of image blocks based on fixed shapes will lead to the splitting of continuous structures, resulting in artifacts during reconstruction. Summary of the invention

[0005] The object of the present invention is to provide a lightweight image super-resolution reconstruction method, device, equipment and medium with super-pixel high-frequency enhancement.

[0006] The technical solution to achieve the purpose of the present invention is: a lightweight image super-resolution reconstruction method with super-pixel high-frequency enhancement,

[0007] The input image is subjected to shallow feature extraction based on a convolutional network, and the extracted shallow features are passed through K superpixel high-frequency enhancement modules SHFEB to extract deep features; the superpixel high-frequency enhancement module SHFEB includes a superpixel attention branch and a high-frequency enhancement branch, wherein the superpixel attention branch includes a superpixel segmentation SPS module, an intra-pixel attention AWP module, and an inter-pixel attention ABP module; for the superpixel attention branch in each SHFEB module, the input features are firstly subjected to superpixel aggregation by the SPS module, and the image is divided into multiple pixel blocks with similar features, namely superpixels, and the superpixels are clustered according to the similarity to obtain the local area;

[0008] The AWP module analyzes and captures the complex dependencies between pixels within each superpixel, and the ABP module captures the interactions between remote pixels. The output features are element-wise added with the high-frequency information extracted by the high-frequency enhancement branch.

[0009] After K SHFEB modules, the output features are reconstructed into a high-resolution image using pixel shuffling.

[0010] Furthermore, the input low-resolution image is first processed through a 3×3 convolutional layer to obtain shallow features.

[0011] Furthermore, for the superpixel attention branch in each SHFEB module, the input features are firstly aggregated by the SPS module to divide the image into multiple pixel blocks with similar features. The specific implementation is as follows:

[0012] The SLIC superpixel segmentation algorithm is used to perform superpixel segmentation on the initial image. For an image with a total of n pixels, the superpixel task is essentially to calculate the association between pixels and superpixels:

[0013] K∈{0,1,...,m-1} n×1

[0014] Where m means that each pixel is assigned to one of the m superpixels;

[0015] First, superpixel centers are sampled by averaging the labels in the shallow feature map, followed by pixel-superpixel association:

[0016]

[0017] Where I p represents the pth pixel, A i represents the cluster center of the i-th superpixel element, K p represents the superpixel element to which the p-th pixel belongs, and the superscript t represents the t-th iteration;

[0018] Finally, the new superpixel cluster center is obtained by calculating the average of all pixel features associated with each superpixel:

[0019]

[0020] In the formula Represents the number of pixels contained in the i-th superpixel element;

[0021] By iterating the above steps until the expectation is maximized, the final association map Q is obtained t .

[0022] Furthermore, the inter-pixel attention module ABP first propagates feature information from pixels to superpixels, and then distributes the aggregated superpixel information to pixels through a cross-attention mechanism; superpixels can generate high-quality super-resolution images by exploiting the complementarity between features; pixel feature y∈R N×C , N represents the number of superpixels, the basic pixel feature x∈RHW×C ; The attention mechanism is used to map the basic pixel information to the superpixel. After completing the feature aggregation in the superpixel, the aggregated information needs to be distributed back to the basic pixel to realize the information propagation between pixels; Get the query Q from the basic pixel features x , use Q y As a key to reduce the number of parameters, the attention mechanism is used again to map the updated superpixel features back to the base pixels; first, the query Q is calculated using linear projection y , key K x Sum value V x , Q x , Q y are all elements required to calculate the attention mechanism.

[0023]

[0024] In the formula and are the weight matrices for query, key, and value respectively; the attention mechanism is expressed as:

[0025]

[0026] Where B is the dynamic relative position code, is the scale factor, y u is the updated superpixel feature.

[0027] Furthermore, the intra-pixel attention module selects the most-N pixels that are most similar to each superpixel and uses the association graph to perform the attention mechanism operation;

[0028] Let z = {x(i)} N ∈R N×C is a basic pixel within a superpixel, where N is the number of selected pixels, and the pixel is updated using the intra-pixel attention mechanism according to the following formula,

[0029]

[0030] The projection weight matrices for query, key, and value are and For basic pixels other than N pixels, use Projection is performed to obtain the updated features, which are finally integrated with the most-N pixels most similar to each superpixel to obtain the final feature information.

[0031] Furthermore, the high-frequency enhancement branch extracts high-frequency information including:

[0032] The high frequency enhancement branch includes a basic feature extraction BFE branch and a high frequency feature extraction HFFE branch;

[0033] First, the input image feature I i ∈R H×W×C It is divided into two parts and sent to the BFE branch and HFFE branch respectively, and the

[0034] The BFE branch uses a 3×3 convolution layer and ReLU activation function to extract local features; the HFFE branch uses a maximum pooling layer to extract high-frequency information from the input features, and then uses a 1×1 convolution layer and ReLU activation function to enhance the high-frequency features;

[0035] Finally, the outputs of the two branches are added pixel by pixel, and the information is thoroughly fused through a 1×1 convolutional layer to obtain the final deep features. At the same time, skip connections are introduced to maintain the stability of the training process.

[0036] A lightweight image super-resolution reconstruction device with super-pixel high-frequency enhancement, comprising:

[0037] A shallow feature extraction module is used to extract shallow features of the input image based on a convolutional network;

[0038] The super-high pixel high-frequency enhancement module is used to extract deep features from the extracted shallow features through N super-pixel high-frequency enhancement modules SHFEB; the super-high pixel high-frequency enhancement module includes a super-pixel attention branch and a high-frequency enhancement branch, wherein the super-pixel attention branch includes a super-pixel segmentation SPS module, an intra-pixel attention AWP module, and an inter-pixel attention ABP module; the input features are first super-pixel aggregated through the SPS module to divide the image into multiple pixel blocks with similar features, namely super-pixels, and the super-pixels are clustered according to the similarity to obtain local areas; the intra-pixel attention module AWP analyzes and captures the complex dependencies between pixels in each super-pixel; the inter-pixel attention module ABP captures the interaction between remote pixels, and the output features obtained are element-wise added with the high-frequency information extracted by the high-frequency enhancement branch;

[0039] The pixel reshuffling module is used to reconstruct the output features into a high-resolution image by using pixel reshuffling.

[0040] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0041] A computer-readable storage medium stores a computer program, which implements the steps of the above method when executed by a processor.

[0042] Compared with the prior art, the present invention has the following significant advantages: (1) The present invention aggregates pixels with similar characteristics to form superpixels. Compared with the image segmentation method using fixed shapes, superpixel segmentation allows adaptive segmentation of the input image, ensuring the continuity of the pixel structure, thereby effectively improving the network's feature extraction capability; and local and global information interaction is achieved through the intra-pixel attention module and the inter-pixel attention module respectively. (2) The high-frequency enhancement module is used to add high-frequency prior information to the network, so that the scheme has a better reconstruction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a network structure diagram of the lightweight image super-resolution reconstruction method with super-pixel high-frequency enhancement of the present invention.

[0044] Figure 2 It is the inter-pixel attention mechanism of the present invention.

[0045] Figure 3 is the intra-pixel attention mechanism of the present invention.

[0046] Figure 4 It is the high frequency enhancement module of the present invention.

[0047] Figure 5 This is the visualization result of different algorithms on the Zrbrn.png image in the Set14 dataset.

[0048] Figure 6 This is the visualization result of different algorithms on Urban100 and BSDS00 datasets. DETAILED DESCRIPTION

[0049] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0050] In view of the problems described in the background technology, the present invention proposes a lightweight image super-resolution reconstruction method based on superpixel high-frequency enhancement. First, a shallow feature extraction based on a convolutional network is performed on the input image, and adjacent pixels are grouped into superpixels through local clustering. The superpixels are clustered according to similarity to obtain local areas. Compared with the image segmentation method using a fixed shape, superpixel segmentation allows the input image to be adaptively segmented to ensure the continuity of the pixel structure. Then, the intra-pixel attention module and the inter-pixel attention module are introduced, and the two attention mechanisms cooperate with each other in local and global feature extraction. At the same time, the high-frequency enhancement branch is used to introduce high-frequency prior information to capture high-frequency details and improve the quality of the reconstructed image.

[0051] The superpixel high-frequency enhanced image super-resolution reconstruction network proposed in the present invention includes a superpixel attention branch and a high-frequency enhancement branch, wherein the superpixel attention branch is composed of a superpixel segmentation (SPS) module, an attention within pixels (AWP) module and an attention between pixels (ABP) module. The overall network structure is as follows: Figure 1 As shown in the figure. The input low-resolution image is first processed by a 3×3 convolutional layer to obtain shallow features, and then passed through K superpixel high-frequency enhancement blocks (SHFEB) to extract deep features. For the superpixel attention branch in each SHFEB block, the input features are first aggregated into superpixels through the SPS superpixel segmentation module. Then, the AWP intra-pixel attention module captures the dependencies of pixels within each superpixel, and the ABP inter-pixel attention module captures the interactions between remote pixels. The output features are element-wise added with the high-frequency information extracted by the high-frequency enhancement branch. After the K SHFEB blocks, pixel-shuffle is used to reconstruct the high-resolution image.

[0052] Superpixel segmentation

[0053] In order to ensure the structural continuity of pixels, the input features are divided into superpixels to obtain more accurate pixel boundaries. This paper adopts the SLIC superpixel segmentation algorithm. For a picture with a total of n pixels, the superpixel task is essentially to calculate the association between pixels and superpixels:

[0054] K∈{0,1,...,m-1} n×1

[0055] Where m means that each pixel is assigned to one of the m superpixels.

[0056] First, superpixel centers are sampled by averaging the labels in a regular grid, followed by pixel-superpixel association:

[0057]

[0058] Where I p represents the pth pixel, A i represents the cluster center of the i-th superpixel element, K p It represents the superpixel element to which the p-th pixel belongs, and the superscript t represents the t-th iteration.

[0059] Finally, the new superpixel cluster center is obtained by calculating the average of all pixel features associated with each superpixel:

[0060]

[0061] In the formula Represents the number of pixels contained in the i-th superpixel element. By iterating the above steps until the expectation is maximized, the final association map Q is obtained. t .

[0062] Pixel-to-Pixel Attention Mechanism

[0063] The inter-pixel attention module ABP first propagates information from pixels to superpixels, and then distributes the aggregated information to pixels through a cross-attention mechanism. Since the features of pixels are highly similar to those of their superpixels, superpixels can exploit the complementarity between features to produce high-quality super-resolution images. Figure 2 As shown, the superpixel feature y∈R N×C , N represents the number of superpixels, the basic pixel feature x∈R HW×C The attention mechanism is used to map the basic pixel information to superpixels. First, the query Q is calculated using linear projection. y , key K x Sum value V x .

[0064]

[0065] In the formula and are the weight matrices for query, key, and value respectively. The attention mechanism can be expressed as:

[0066]

[0067] Where B is the dynamic relative position code, is the scale factor, y u is the updated superpixel feature.

[0068] After completing feature aggregation in superpixels, the aggregated information needs to be distributed back to the basic pixels to achieve information propagation between pixels. Get the query Q from the basic pixel features x , use Q y As a key to reduce the number of parameters, the attention mechanism is again utilized to map the updated superpixel features back to the underlying pixels.

[0069] In-pixel attention mechanism

[0070] Using the complementarity of similar pixels in the same superpixel helps improve the quality of super-resolution images, but some superpixels contain a large number of pixels, which increases the complexity of the model. Therefore, the intra-pixel attention mechanism selects the most-N pixels that are most similar to each superpixel and uses the association graph Q t Perform attention mechanism operation. Figure 3 As shown, let z = {x(i)} N ∈R N×C is the basic pixel within a superpixel, where N is the number of selected pixels, and intra-pixel attention is performed according to Formula 5 and Formula 6, where the projection weight matrices of query, key and value are and For basic pixels other than N pixels, use Projection is performed to obtain updated features, which are finally integrated with the updated pixels within the superpixel to obtain the final feature information.

[0071] High frequency enhancement module

[0072] like Figure 4 As shown in FIG. 1 , the high frequency enhancement module includes a basic feature extraction (BFE) branch and a high frequency feature extraction (HFFE) branch. First, the input feature I i ∈R H×W×C It is divided into two parts and sent to the BFE branch and HFFE branch respectively, and the The BFE branch uses a 3×3 convolution layer and ReLU activation function to extract local features; the HFFE branch uses a maximum pooling layer to extract high-frequency information in the input features, followed by a 1×1 convolution layer and ReLU activation function to enhance the high-frequency features; finally, the outputs of the two branches are added pixel by pixel, and the information is thoroughly fused through a 1×1 convolution layer, while skip connections are introduced to maintain the stability of the training process.

[0073] Experimental data and indicators

[0074] The method of the present invention uses DIV2K as a training dataset, which includes 900 high-resolution images of natural scenes, of which 800 images are used for model training and 100 images are used for model verification. Data expansion is performed by randomly rotating and horizontally flipping the 800 training images. The LR samples in the training phase are generated by the bicubic downsampling method. In the test phase, four standard datasets, Set5, Set14, BSDS100 and Urban100, are used, with magnifications of ×2, ×3, and ×4 respectively, and PSNR and SSIM are used to evaluate the quality of the reconstructed images.

[0075] Experimental environment and settings

[0076] The platform and main parameters used in the experiment of this invention: NVIDIA GeForce RTX 3060, video memory 12G; system environment: Linux; programming language: Python; deep learning framework: Pytorch.

[0077] The initial learning rate in training is set to 0.0002, the output channels of all modules are set to 40, and the number of SHFEB blocks is 8. Quantitative comparison is done on the Y channel by converting RGB format images to YCbCr format.

[0078] Comparative experimental results and analysis

[0079] In order to objectively evaluate the reconstruction performance of the proposed algorithm, the proposed method is compared with a variety of lightweight SR models and transformer-based models, including DRSAN, SMSR, CARN, etc. The following analysis is conducted from two aspects: quantitative indicators and visualization results.

[0080] (1) Quantitative analysis: The average PSNR and SSIM values ​​of different algorithms on five test sets are shown in Table 1. Under the three magnification factors, the PSNR and SSIM values ​​of the proposed method are improved compared with those of other algorithms. At the same time, it can be seen that the proposed model performs better in the Urban100 dataset that contains more details.

[0081] Table 1 Average PSNR and SSIM of different algorithms on four datasets

[0082]

[0083]

[0084] (2) Visual analysis: Figure 5 The comparison of the model generation effect when the magnification is 4 is shown. It can be seen that the image obtained by bicubic interpolation is blurred overall, and the reconstructed images obtained by other methods are also partially blurred. The proposed algorithm can restore high-frequency details to the greatest extent.

[0085] Figure 6 The reconstruction effects of two images in the BSDS100 and Urban100 datasets are shown when the magnification is 4. It can be seen that other algorithms will produce incorrect reconstruction of the direction of texture details in the image, while the method proposed in the present invention can better restore the edge and texture details of the image, and the reconstructed image is closer to the real image.

[0086] The lightweight image super-resolution reconstruction method based on superpixel high-frequency enhancement proposed in this paper uses superpixels to group local similar pixels into interpretable local areas; adopts the intra-pixel attention mechanism to promote local information interaction in irregular local superpixel areas; and the inter-pixel attention module promotes remote information interaction by replacing superpixels; and uses the high-frequency enhancement module to enhance high-frequency information. Experimental results show that the method proposed in this paper achieves better reconstruction performance than other mainstream super-resolution reconstruction algorithms.

Claims

1. A lightweight image super-resolution reconstruction method with super-pixel high-frequency enhancement, characterized in that: The input image is subjected to shallow feature extraction based on a convolutional network, and the extracted shallow features are passed through K super-pixel high-frequency enhancement modules SHFEB to extract deep features; The superpixel high-frequency enhancement module SHFEB includes a superpixel attention branch and a high-frequency enhancement branch, wherein the superpixel attention branch includes a superpixel segmentation SPS module, an intra-pixel attention AWP module, and an inter-pixel attention ABP module; for the superpixel attention branch in each SHFEB module, the input features are firstly aggregated by the SPS module for superpixels, and the image is divided into multiple pixel blocks with similar features, namely superpixels, and the superpixels are clustered according to the similarity to obtain the local area; The AWP module analyzes and captures the complex dependencies between pixels within each superpixel, and the ABP module captures the interactions between remote pixels. The output features are element-wise added with the high-frequency information extracted by the high-frequency enhancement branch. After K SHFEB modules, the output features are reconstructed into a high-resolution image using pixel shuffling.

2. The lightweight image super-resolution reconstruction method with super-pixel high-frequency enhancement according to claim 1, characterized in that: The input low-resolution image is first processed through a 3×3 convolutional layer to obtain shallow features.

3. The lightweight image super-resolution reconstruction method with super-pixel high-frequency enhancement according to claim 1, characterized in that: For the superpixel attention branch in each SHFEB module, the input features are first aggregated by the SPS module to divide the image into multiple pixel blocks with similar features. The specific implementation is as follows: The SLIC superpixel segmentation algorithm is used to perform superpixel segmentation on the initial image. For an image with a total of n pixels, the superpixel task is essentially to calculate the association between pixels and superpixels: K∈{0,1,...,m-1} n×1 Where m means that each pixel is assigned to one of the m superpixels; First, superpixel centers are sampled by averaging the labels in the shallow feature map, followed by pixel-superpixel association: Where I p represents the pth pixel, A i represents the cluster center of the i-th superpixel element, K p represents the superpixel element to which the p-th pixel belongs, and the superscript t represents the t-th iteration; Finally, the new superpixel cluster center is obtained by calculating the average of all pixel features associated with each superpixel: In the formula Represents the number of pixels contained in the i-th superpixel element; By iterating the above steps until the expectation is maximized, the final association map Q is obtained t .

4. The lightweight image super-resolution reconstruction method with super-pixel high-frequency enhancement according to claim 1, characterized in that: The inter-pixel attention module ABP first propagates feature information from pixels to superpixels, and then distributes the aggregated superpixel information to pixels through a cross-attention mechanism; superpixels can produce high-quality super-resolution images by exploiting the complementarity between features; Pixel feature y∈R N×C , N represents the number of superpixels, the basic pixel feature x∈R HW×C ; The attention mechanism is used to map the basic pixel information to the superpixel. After completing the feature aggregation in the superpixel, the aggregated information needs to be distributed back to the basic pixel to realize the information propagation between pixels; Get the query Q from the basic pixel features x , use Q y As a key to reduce the number of parameters, the attention mechanism is used again to map the updated superpixel features back to the base pixels; first, the query Q is calculated using linear projection y , key K x Sum value V x , Q x , Q y are all elements required to calculate the attention mechanism. In the formula and are the weight matrices for query, key, and value respectively; the attention mechanism is expressed as: Where B is the dynamic relative position code, is the scale factor, y u is the updated superpixel feature.

5. The lightweight image super-resolution reconstruction method with super-pixel high-frequency enhancement according to claim 1, characterized in that: The intra-pixel attention module selects the most-N pixels that are most similar to each superpixel and uses the association graph to perform the attention mechanism operation; Let z = {x(i)} N ∈R N×C is a basic pixel within a superpixel, where N is the number of selected pixels, and the pixel is updated using the intra-pixel attention mechanism according to the following formula, The projection weight matrices for query, key, and value are and For basic pixels other than N pixels, use Projection is performed to obtain the updated features, which are finally integrated with the most-N pixels most similar to each superpixel to obtain the final feature information.

6. The lightweight image super-resolution reconstruction method with super-pixel high-frequency enhancement according to claim 1, characterized in that: The high-frequency enhancement branch extracts high-frequency information including: The high frequency enhancement branch includes a basic feature extraction BFE branch and a high frequency feature extraction HFFE branch; First, the input image feature I i ∈R H×W×C It is divided into two parts and sent to the BFE branch and HFFE branch respectively, and the The BFE branch uses a 3×3 convolution layer and ReLU activation function to extract local features; the HFFE branch uses a maximum pooling layer to extract high-frequency information from the input features, and then uses a 1×1 convolution layer and ReLU activation function to enhance the high-frequency features; Finally, the outputs of the two branches are added pixel by pixel, and the information is thoroughly fused through a 1×1 convolutional layer to obtain the final deep features. At the same time, skip connections are introduced to maintain the stability of the training process.

7. A lightweight image super-resolution reconstruction device with super-pixel high-frequency enhancement, characterized in that: include: A shallow feature extraction module is used to extract shallow features of the input image based on a convolutional network; The super-pixel high-frequency enhancement module is used to extract the deep-level features by passing the extracted shallow features through N super-pixel high-frequency enhancement modules SHFEB; The super-pixel high-frequency enhancement module includes a super-pixel attention branch and a high-frequency enhancement branch, wherein the super-pixel attention branch includes a super-pixel segmentation SPS module, an intra-pixel attention AWP module, and an inter-pixel attention ABP module; The input features are first aggregated by the SPS module to perform superpixel aggregation, dividing the image into multiple pixel blocks with similar features, namely superpixels. The superpixels are clustered according to their similarity to obtain local areas. The intra-pixel attention module AWP analyzes and captures the complex dependencies between pixels within each superpixel. The inter-pixel attention module ABP captures the interactions between remote pixels, and the output features are element-wise added with the high-frequency information extracted by the high-frequency enhancement branch. The pixel reshuffling module is used to reconstruct the output features into a high-resolution image by using pixel reshuffling.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images